--- title: PST Hierarchy in Multi-Domain Research url: https://www.emergentmind.com/topics/pst-hierarchy type: topic --- # PST Hierarchy in Multi-Domain Research Across the cited literature, “PST hierarchy” does not denote a single canonical object. It names several unrelated hierarchical constructions: a strict parameter-age segmentation of a single network in continual learning [1905.11550], the subgroup-property framework that characterizes soluble PST-groups in finite group theory [1304.7902], a rooted-tree enrichment of zero-dimensional persistence pairs [1907.13495], an intermediate semidefinite-programming hierarchy for entanglement detection [2508.05854], and Pearl’s three-layer observational–interventional–counterfactual hierarchy as formalized in potential-outcomes notation [2601.20405]. What unifies these usages is not a shared formalism but the role of hierarchy itself: each framework imposes an order relation that structures inference, computation, or classification. ## 1. Terminological scope The abbreviation PST is polysemous. In the materials considered here, it refers to hierarchies over network parameters, subgroup properties, persistence pairs, semidefinite relaxations, and causal estimands. The resulting objects differ in algebraic type, but each organizes a family of entities by a monotone relation. | Domain | PST hierarchy | Ordering principle | |---|---|---| | Continual learning | Progressive segmented training | Older task segments are frozen; newer tasks use remaining capacity | | Finite group theory | PST-group characterizations | Stronger and weaker subgroup properties; Hall-type criteria | | Topological data analysis | Rooted tree on persistence pairs | Ancestry induced by nested branch-merges | | Quantum information | Partial-PPT extension hierarchy | $EXT_k \supseteq PST_k \supseteq DPS_k$ | | Causal inference | P–S–T hierarchy | Observational $\to$ interventional $\to$ counterfactual | A common source of confusion is to treat these as variants of one theory. They are not. The continual-learning usage concerns parameter allocation in a neural network; the group-theoretic usage concerns transitivity of $S$-permutability; the topological usage augments persistence diagrams with branch structure; the quantum-information usage defines a cone between EXT and DPS; and the causal-inference usage stratifies estimands by the type of potential-outcomes object required. ## 2. Progressive segmented training as a hierarchy of network capacity In continual learning, PST denotes “Progressive Segmented Training,” a single-network method that builds a strict hierarchy of non-overlapping parameter segments across tasks [1905.11550]. Just before learning task $T_i$, the parameter set is decomposed as $$ \Theta = (\Theta_{\text{fixed}}, \Theta_{\text{free}}), $$ where $\Theta_{\text{fixed}}$ is frozen to preserve previously learned tasks and $\Theta_{\text{free}}$ is available for adaptation to $T_i$. Training uses new data $X^{(i)}$ together with a fixed-size memory $P$ containing up to $K$ examples from previous tasks. The training loss is a memory-assisted classification loss on a balanced batch drawn equally from $X^{(i)}$ and $P$, with only $\Theta_{\text{free}}$ updated. After training the free parameters on task $T_i$, PST scores each elementary unit by a first-order Taylor criterion. For a convolutional filter, $$ s_\ell^o \simeq \sum_{i,m,n}\left|\frac{\partial L}{\partial \Theta_\ell^{o,i,m,n}}\cdot \Theta_\ell^{o,i,m,n}\right|, $$ and for a fully connected neuron, $$ s_\ell^t \simeq \sum_i\left|\frac{\partial L}{\partial \Theta_\ell^{t,i}}\cdot \Theta_\ell^{t,i}\right|. $$ Within each layer, units are ranked by score, and the top $\beta_\ell\%$ are selected as $\Theta_{\text{important}}$; the remainder is $\Theta_{\text{secondary}}$. The hierarchy is produced by the segmented update rules. PST re-initializes the selected important weights, retrains only those weights for a small number of epochs, freezes the reinforced result by moving it into $\Theta_{\text{fixed}}$, and releases $\Theta_{\text{secondary}}$ as the new free pool for the next task. Over time this yields $$ \Theta = (\Theta_{\text{fixed}}^{(i)}, \Theta_{\text{free}}^{(i+1)}), $$ with $\Theta_{\text{fixed}}^{(i)}$ containing all parameters important to tasks $T_1,\dots,T_i$. The supplied description states that this produces a strict hierarchy by task age: older tasks sit deeper in the frozen portion of the network, while newer tasks progressively claim a small slice of the capacity. The fixed-size memory buffer is explicitly part of the stability–plasticity mechanism. Memory is uniformly sampled over previously seen classes, and replay is injected in three stages: initial epochs, periodic replay, and final classifier fine-tuning. This design is used without additional regularization. Empirically, the method achieves state-of-the-art accuracy in the single-head evaluation on CIFAR-10 and CIFAR-100, and the shrinking free pool yields a computational advantage: on CIFAR-100, PST reports a $>24\times$ reduction in weight-update FLOPs and an $\approx 1.5\times$ reduction in total train-time FLOPs by the final tasks. The important conceptual point is that the hierarchy is not architectural growth; it is progressive freezing inside a single network. ## 3. The PST-group hierarchy in finite group theory In finite group theory, a PST-group is a finite group in which $S$-permutability is transitive [1304.7902]. Formally, if $H\le K\le G$, with $H$ $S$-permutable in $K$ and $K$ $S$-permutable in $G$, then $H$ is $S$-permutable in $G$. The classical characterization cited in the source states that a finite soluble group $G$ is a PST-group if and only if $$ G=D\rtimes M, $$ where $D=G^{\mathfrak N}$ is an abelian Hall subgroup of $G$, and every $m\in M$ induces on $D$ only power automorphisms. In particular, $D\le Z(N_G(D))$ and $G$ splits over $D$. The 2013 paper refines this theory via quasipermutability and $S$-quasipermutability. If $G=N_G(H)B$, then $H$ is quasipermutable if it permutes with $B$ and with every subgroup $A\le B$ such that $(|H|,|A|)=1$. It is $S$-quasipermutable if it permutes with $B$ and with every Sylow subgroup $P\le B$ such that $(|H|,|P|)=1$. These notions sit strictly below classical permutability conditions: requiring permutation with all subgroups recovers quasinormality, while requiring permutation with all Sylow subgroups recovers $S$-permutability. The principal Hall-type characterization is Theorem A. Let $D=G^{\mathfrak N}$ and $T=\pi(D)$. Then the following are equivalent: $D$ is a Hall subgroup and every Hall subgroup of $G$ is quasipermutable; $G$ is a soluble PST-group; every subgroup of $G$ is quasipermutable; and every $T$-subgroup together with some minimal supplement of $D$ is quasipermutable. This equivalence is the key sense in which hierarchy enters the theory: a Hall-level criterion already determines the global PST property. The paper also places PST-groups inside a broader hierarchy of subgroup properties: $$ \text{semi-normal} \implies \text{propermutable} \implies \text{quasipermutable}, $$ with the parallel chain $$ \text{SS-quasinormal} \implies \text{$S$-propermutable} \implies \text{$S$-quasipermutable}. $$ On the global scale, the text records $$ \text{PT-group} \Downarrow \text{PST-group} \Uparrow \text{SQT-group}. $$ A common misconception is to conflate quasipermutability with quasinormality or $S$-permutability. The source explicitly distinguishes them: the “quasi” condition is relative to a complement $B$ of $N_G(H)$ and only concerns coprime-order subgroups or Sylow subgroups of $B$. ## 4. The rooted-tree hierarchy on persistence pairs In topological data analysis, the PST hierarchy is a rooted-tree structure on zero-dimensional persistence pairs, introduced to capture spatial relations that ordinary persistence diagrams do not encode [1907.13495]. The paper itself calls the construction the Interlevel-Set Persistence Hierarchy, and the supplied exposition identifies it as a zero-dimensional Persistence-Space-Tree hierarchy. Its motivation is the absence of spatial relationships between features in persistence diagrams, which limits expressive power. Let $f\colon X\to\mathbb{R}$ be a Morse-type function with distinct critical values, and let $$ X_c=\{x\in X\mid f(x)\le c\}. $$ A zero-dimensional persistence pair is a pair of critical points $(\sigma,\tau)$ where $\sigma$ is a local minimum with birth time $b=f(\sigma)$ and $\tau$ is the critical point at which that component merges into an older one, with death time $d=f(\tau)$. The persistence is $$ \operatorname{pers}(\sigma,\tau)=d-b. $$ The global minimum yields a root pair $(b,\infty)$. The hierarchy places one node for each pair $p=(b,d)$ and defines parent–child relations through branch-merges. The supplied definition uses the condition that two birth points remain in distinct connected components of an interlevel set $$ X_{[\min(b,b'),d')}=\{x\in X\mid \min(b,b')\le f(x)\le d'\}, $$ which distinguishes a genuine branch-merge from a mere extension of an existing branch. The induced order $$ p\le_{\rm PST} q \Longleftrightarrow \text{there is a directed path from } q \text{ down to } p $$ makes the hierarchy a rooted tree with exactly one root and no node with more than one parent. Algorithmically, the construction specializes union-find to persistence computation. Critical events are processed in ascending order, components are labeled by their highest minima, and when components merge one tests whether the corresponding minima remain separated in the interlevel set. If they do, a child edge is added in the hierarchy. This augments the usual persistence pairing with branching information. Two derived quantities make the structure useful beyond visualization. First, the rank of a node $v$ is the number of descendants: $$ \operatorname{rank}(v)=\bigl|\{w\in H\mid v\le_{\rm PST} w\}\bigr|. $$ Second, the supplied exposition defines edge and node stability measures based on the birth and death coordinates of adjacent pairs. These quantities are intended to measure how stable the pairing and its branch structure are under perturbation. The conceptual significance is that the hierarchy records which features are nested inside others, thereby encoding branching structure that plain persistence diagrams omit. ## 5. The partial-PPT extension hierarchy for entanglement detection In quantum information, PST denotes a new semidefinite-programming hierarchy for entanglement detection, positioned strictly between the EXT and DPS hierarchies [2508.05854]. If $\mathcal{S}_{\rm ep}$ denotes the cone of separable operators on a bipartite space $\mathcal{H}_a\otimes\mathcal{H}_b$, the hierarchy is defined through symmetric extensions: $$ EXT_k := \{\rho_{ab}\mid \text{there exists a $k$-party symmetric extension of } \rho_{ab}\}, $$ $$ DPS_k := \{\rho_{ab}\mid \rho_{ab} \text{ admits a $k$-party symmetric extension that is PPT on all subsystems}\}. $$ The intermediate cone is $$ PST_k := \left\{\rho_{ab}\ \middle|\ \exists\, \rho_{aB}\succeq 0,\ \operatorname{Tr}_{b_2\dots b_k}\rho_{aB}=\rho_{ab},\ \rho_{aB}\ \text{symmetric under }S_k,\ (\rho_{aB})^{T_{b_1\dots b_k}}\succeq 0\right\}. $$ Hence, for each $k$, $$ EXT_k \supseteq PST_k \supseteq DPS_k, $$ and in the limit, $$ \bigcap_k EXT_k=\bigcap_k PST_k=SEP. $$ A central contribution is a polynomial-size description of EXT and PST via a partition matrix $\Pi$ onto the symmetric subspace and a lifting-down operator $\Lambda$. The action of $\Lambda$ and its adjoint can be implemented in $O(d\cdot d_k)$ flops rather than in dimension $d^k$, which is the key compression step. The paper states that these compact descriptions satisfy the Slater condition for both primal and dual SDPs, ensuring strong duality and the absence of duality gaps. The hierarchy is accompanied by tailored algorithms. Three first-order methods are developed from least-squares formulations—Frank–Wolfe, projected gradient, and fast projected gradient—and a custom primal–dual interior-point method is derived from a conic formulation. The complexity statement is explicit. If $n=d_a\cdot d_k$ and $m=d_a^2\cdot d_b^2$, each primal–dual interior-point iteration for EXT$_k$ or PST$_k$ costs $$ O(m^2 n^2 + m n^3 + n^3) $$ flops and converges in $$ O(\sqrt{n}\log(1/\epsilon)) $$ iterations for accuracy $\epsilon$. Each first-order iteration costs $$ O(m n^2) $$ flops and yields a duality gap of $O(1/T)$ or $O(1/T^2)$ after $T$ iterations. The numerical position of PST is also specific. The paper reports that EXT$_1$ or EXT$_2$ with first-order methods quickly detects strongly entangled states, while PST$_2$ plus first-order methods extends detection to weaker entanglement. The implementations scale to levels $k\approx 15$ in minutes, compared with previous limits $k\approx 3$–$4$, and the custom interior-point method is faster than off-the-shelf SDP solvers such as MOSEK via PICOS. The essential correction to a possible misunderstanding is that PST is neither a replacement for EXT nor for DPS; it is an intermediate cone that is tighter than EXT and cheaper than DPS. ## 6. The P–S–T causal hierarchy In causal inference, PST refers to Pearl’s three-layer causal hierarchy, recast in potential-outcomes notation as P, S, and T [2601.20405]. The layers are defined by the class of objects required to express a query. Let $A$ denote treatment, $X$ pre-treatment covariates, and $Y$ the outcome, with potential outcomes $Y(a)$ and consistency $Y=Y(A)$. Layer P is observational or “seeing.” It contains only the joint and conditional distributions of observed variables: $$ P(X),\quad P(A,X),\quad P(Y,X,A),\quad P(Y\mid X=x),\quad P(A\mid X=x). $$ Queries at this level remain within the observed $\sigma$-algebra, such as $E[Y\mid A=a,X=x]$. Layer S is interventional or “doing.” It contains marginal distributions of potential outcomes under single interventions: $$ P(Y=y\mid do(A=a),X=x)\equiv P(Y(a)=y\mid X=x), $$ and $$ P(Y(a)=y)=\sum_x P(Y(a)=y\mid X=x)P(X=x). $$ Typical estimands include the average treatment effect $E[Y(1)-Y(0)]$, the conditional ATE, quantile treatment effects, dose–response functions, and the controlled direct effect $E[Y(1,m)-Y(0,m)]$. Layer T is counterfactual or “imagining.” It contains joint or cross-world features of potential outcomes, nested counterfactuals, and individual-level contrasts: $$ (Y(0),Y(1)),\qquad Y(a,M(a')), \qquad Y_i(1)-Y_i(0). $$ Representative estimands include the probability of necessary causation, probability of sufficient causation, treatment benefit and harm rates, persuasion rate, natural direct and indirect effects, principal causal effects, and the individual treatment effect. An important technical distinction in the source is that controlled direct effects lie in layer S, whereas natural direct and indirect effects lie in layer T because they involve nested counterfactuals. The hierarchy is also a hierarchy of identifiability difficulty. For layer P, no causal assumptions are needed. For layer S, the principal obstacle is confounding, and the standard assumptions are ignorability, consistency, and overlap: $$ \{Y(0),Y(1)\}\perp A\mid X,\qquad Y=Y(A),\qquad 0